We know a great deal about the content of children's theories of the physical world: from the earliest ages, they represent objects as discrete, impenetrable entities that cannot wink in and out of existence and provoke motion only on contact. Computational work has crystallized these insights into simulation theories to striking success, describing how children represent physical scenes and make predictions based on them. Yet we know that having a simulator is not enough: there's so much people can do that, if people are to do anything, they must leverage high-level concepts to direct their actions. The question of how this is done has gotten far less attention than it deserves, as there's too much action for people who like concepts and too much concepts for people who like action. In this talk, we'll investigate the form that these concepts can take such that they might guide simulations, as well as how they can be learned. In particular, we'll focus on tools as a royal road to studying how conceptual structures inform action.